Learning in MultiAgent Systems Essay

2.6 Learning in MultiAgent Systems
MultiAgent Systems have been commonly used to apply ideas of decentralization in complex problems. These ideas include the fields of distributed systems, e.g. Actor model, as described earlier in Section 2.4. In distributed systems a number of entities cooperate to achieve solutions. The application of artificial intelligence within distributed systems is referred to as distributed artificial intelligence or DAI. Two types of approaches to DAI are usually identified, where first one is focused on the distributed problem solving and deals with the subdivision of the problem solving between the nodes of the network. The second type is focused on the collective behaviors of the problem solving agents within the multiagent system (MAS). The agents have a certain level of autotomy and the approach focuses on studying the complex interactions between the agents. The latter approach directly relates to the definition of the decentralized autonomous organizations (DAO) from Section 2.5.
Figure 24 Learning Agent by Atmaram [49]
The definition of agent is provided by Panait & Luke [50] as an autonomous computational entity which receives information from its environment and performs actions based on this information. In the multiagent environment several agents interact with each other and the environment also poses certain constraints on agents and their behavior, e.g. agents may not have complete information about the entire environment and all the other agents. This property of MAS is important in using MAS for problem solving and prevents the agents from knowing everything and acting synchronously, which can then reduce the problem to be solvable by a single agent. According to Jennings et al. [51], few complex problems in the real world allow such simplifications, as most problems are characterized by incomplete agents having information about the world, decentralized decisionmaking, and asynchronous and distributed computation.
Existing literature distinguishes several application classifications of MAS. Implementations of MAS in swarm robotics have been studied [52] based on different properties of agents and agent teams, e.g. processing throughput, team size, and communication ranges. Others studies focused on the applications of MAS in the industries [53] [54] with two main approaches focusing on agents and on the system itself. Agent properties studied include heterogeneity of teams, computational and communication abilities, while in systems approach studies focus on environments settings, e.g. agent communication and interaction rules.
Learning in MAS falls under the area of machine learning. Different approaches in implementing machine learning in MAS exist, e.g. genetic programming, reinforcement learning, Qlearning. The learning in MAS can also be classified as multiagent learning and singleagent learning. In a singleagent learning environment only a single agent can learn while the actions and behaviors of other agents remain static. On the other hand, in a multiagent learning environment agents behaviors and actions are dynamic since all the agents are able to learn and the context is also constantly changing. In this setting learning triggers behavioral changes and even small behavioral changes may have unpredictable effects on the agents and the system on the macrolevel [50].